By David Stephen
Sedona, AZ — It is unlikely that AI regulation would be fully achieved from outside in, meaning physical decrees to the digital sphere, without digital bots—monitoring and enforcing, at least in common areas of the internet.
There are several tips on how to spot AI fakes, precautions against AI audio imitations, misinformation awareness, and so forth, but these are very pedestrian from the fight, at the source, that should be taking place against those outputs.
Major tech companies have announced labeling for AI images and videos, others have banned them from political ad campaigns, there are deepfake detectors that work great for particular AI models that made them, there is the Coalition for Content Provenance and Authenticity, C2PA and efforts by the National Association of Secretaries of State (NASS).
The Arizona Secretary of State, Adrian Fontes, recently made a deepfake video of himself to show voters what is possible. These are great, just that without AI-on-AI detention, efforts may falter.
The US AI Safety Institute and UK’s—with a new office in San Francisco—would eventually have to develop technical capabilities to monitor public AI models as well as their outputs against usefulness for harm. This will be an alert and takedown technical agent—or bot
Large language models (LLMs) are, in part, mathematical frameworks of data, ranging from linear regression to Bayesian probability, calculus and several others. LLMs have shown potency in many directions, yet they make errors—or hallucinate, they also have sub-intent, and often lack some understanding of the world.
An important research for AI Safety Institutes would be to develop ways to apply new combinatorial functions and limits, for monitoring of AIs and their outputs, especially if there is a labeling standard for those to be used in jurisdictions.
Simply, for an AI model or its output to be used in a jurisdiction—state or nationally—it would need to have some kind of ID, where its outputs can be traced, especially if they are used for harm. Those without the labels may not be allowed, just like an airspace or radio frequency identification.
Even without the labeling, it could be possible to use (AI) monitoring to understand the aims of the output in the scenario—fast enough, for freedom but within order. There are several math functions to explore for this possibility like modular transformations and others.
AI safety and alignment also have to be standardized by the human mind and the human society. The human mind has intentionality for some functions. Context windows of LLMs may need to include new matrix rows for intent, against usefulness for harm, as well as against bias and discrimination.
This intent would also be capped, say with some cost function, such that it does not develop into a destructive one, wedging against some of the existential risks of AI. The human mind, conceptually, mechanizes intentionality or free will in the same sets of electrical and chemical signals for functions. This could be adapted for alignment.
